adaptive-memory-multi-model-router 2.13.26 → 2.14.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (10) hide show
  1. package/README.md +3 -3
  2. package/README_ja.md +75 -11
  3. package/README_zh.md +71 -30
  4. package/_schema.html +19 -46
  5. package/docs/_config.yml +5 -5
  6. package/docs/index.html +3 -3
  7. package/docs/sitemap.xml +24 -6
  8. package/llms.txt +31 -50
  9. package/package.json +2 -2
  10. package/scripts/push-to-gitee.sh +17 -44
package/README.md CHANGED
@@ -1,8 +1,8 @@
1
1
  [🇨🇳 中文](./README_zh.md) · [🇯🇵 日本語](./README_ja.md) · [English](./README.md)
2
2
 
3
- # A3M Router 🔀 — #1 LLM Routing Benchmark # A3M Router 🔀 — 🥇 #1 on RouterArena · Cheapest at $0.047/1K Queries Cheapest Router with Memory
3
+ # A3M Router 🔀 — #1 LLM Routing Benchmark & Cheapest Router with Memory
4
4
 
5
- **#1 on the LLM routing benchmark The highest-ranked and lowest-cost LLM router on the [RouterArena leaderboard](https://github.com/RouteWorks/RouterArena/pull/113). cheapest router with memory. Parallel multi-LLM execution across 47+ providers.**
5
+ **#1 on the LLM routing benchmark & cheapest router with memory.** Parallel multi-LLM execution across 47+ providers.
6
6
 
7
7
  [![npm](https://img.shields.io/npm/dt/adaptive-memory-multi-model-router?color=blue&label=weekly%20downloads)](https://www.npmjs.com/package/adaptive-memory-multi-model-router)
8
8
  [![npm](https://img.shields.io/npm/v/adaptive-memory-multi-model-router)](https://www.npmjs.com/package/adaptive-memory-multi-model-router)
@@ -16,7 +16,7 @@
16
16
 
17
17
  [![npm Downloads Chart](./docs/npm-downloads-chart.svg)](https://www.npmjs.com/package/adaptive-memory-multi-model-router)
18
18
 
19
- > **#1 on LLM routing benchmark > **The highest-ranked and lowest-cost LLM router on the RouterArena leaderboard** — #1 (76.43), cheapest at $0.047/1K queries. Parallel multi-LLM execution across 47+ providers. Open-source, 19.5KB, zero ML dependencies. cheapest router with memory** — RouterArena score 76.43, cost $0.047/1K queries. 47+ providers, 19.5KB, zero ML deps.
19
+ > **#1 LLM routing benchmark & cheapest router with memory** — RouterArena score 76.43, cost $0.047/1K queries. 47+ providers, 19.5KB, zero ML dependencies.
20
20
 
21
21
  **One prompt in. The right model out.** An open-source **AI gateway** that auto-routes every query to the cheapest capable model across **47+ LLM providers**. Features **parallel ensemble execution**, **semantic cache**, **budget enforcement**, **intelligent failover**, and **independent benchmark validation**. Start in <100ms. Python SDK + TypeScript SDK.
22
22
 
package/README_ja.md CHANGED
@@ -1,8 +1,27 @@
1
- # A3M Router 🔀
1
+ # A3M Router 🔀 — LLMルーティングベンチマーク#1 & 最安値メモリ付きルーター
2
2
 
3
- **Intelligent LLMルーティング · 99.5%精度 · ML不要 · GPU不要 · 47プロバイダー**
3
+ **🏆 RouterArenaベンチマーク#1 (76.43) · 最安値 $0.047/1Kリクエスト · 47+プロバイダー並列実行**
4
4
 
5
- マルチシグナル複雑度スコアを使用して99.5%のルーティング精度を実現し、機械学習なしで動作します。各クエリを最も安価な利用可能なモデルに自動ルートします。
5
+ [English](./README.md) | [中文](./README_zh.md) | [日本語](./README_ja.md)
6
+
7
+ ## 主要メトリクス
8
+
9
+ | メトリクス | A3M Router | Sqwish | Azure (Microsoft) | GPT-5 (OpenAI) | RouteLLM (Berkeley) |
10
+ |------------|:----------:|:------:|:------------------:|:---------------:|:-------------------:|
11
+ | **ランキング** | **🏆 #1** | #2 | #3 | #4 | #5 |
12
+ | **スコア** | **76.43** | 75.27 | 71.87 | 64.32 | 48.07 |
13
+ | **コスト** | **$0.047** | $0.18 | $0.22 | $10.02 | $0.27 |
14
+
15
+ > RouterArena公式ベンチマークで最高スコアかつ最低コストを達成（独立評価パイプライン検証 arXiv:2510.00202）
16
+
17
+ ## 独自機能：並列マルチLLM実行
18
+
19
+ 従来のルーターは1つずつモデルを試します（シーケンシャルフォールバック）。A3Mは**複数プロバイダーを並列実行**し、信頼度スコアリングで最良の結果を選択します。
20
+
21
+ ```
22
+ 従来: モデルA ❌ → モデルB ❌ → モデルC ✅ (3倍レイテンシ)
23
+ A3M: モデルA ║ モデルB ║ モデルC → スコアリングで最良を選択 ✅ (1倍レイテンシ)
24
+ ```
6
25
 
7
26
  ## 対応中国LLM
8
27
 
@@ -15,22 +34,67 @@
15
34
  | **Yi (01.AI)** | Yi-1.5, 34B | バイリンガル+長コンテキスト |
16
35
  | **MiniMax** | abab6.5 | 1Mコンテキスト |
17
36
 
37
+ ## コア機能
38
+
39
+ - 🏆 **RouterArena #1** — 19ルーター中1位
40
+ - 🔀 **並列マルチLLM実行** — 複数プロバイダー同時実行、信頼度投票
41
+ - 💰 **最安値** — $0.047/1Kリクエスト、#2より4倍安い
42
+ - 🧠 **メモリ付きルーティング** — エピソードック記憶でセッション越えコンテキスト保存
43
+ - 🔄 **セマンティックキャッシュ** — 30%+ヒット率、コスト節約
44
+ - 🛡️ **予算強制** — クエリごとコスト追跡、超過防止
45
+ - ⚡ **高速起動** — <100ms、19.5KB、ML依存なし
46
+ - 🌐 **47+プロバイダー** — OpenAI, Anthropic, DeepSeek, Groq, NVIDIA等
47
+
18
48
  ## クイックスタート
19
49
 
20
50
  ```bash
21
51
  npm install adaptive-memory-multi-model-router
22
- npx a3m-router serve
52
+ npx a3m-router route "Pythonでソート関数を書いて"
23
53
  ```
24
54
 
25
- OpenAI SDKを`http://localhost:8787/v1`に向けるだけで、コード変更不要。
55
+ ```javascript
56
+ import { createRouter } from 'adaptive-memory-multi-model-router';
57
+
58
+ const router = createRouter({ /* APIキー自動検出 */ });
59
+ const result = await router.route('量子コンピューティングを説明して');
60
+ console.log(result.response); // AI応答
61
+ console.log(result.provider); // 使用プロバイダー
62
+ console.log(result.cost); // 実際のコスト
63
+ ```
64
+
65
+ ## メモリ機能
66
+
67
+ ```javascript
68
+ const router = createRouter({
69
+ memory: { enabled: true } // 会話コンテキスト自動保存
70
+ });
71
+
72
+ // 1回目の会話
73
+ await router.route('私の名前は太郎です'); // 記憶：ユーザー名＝太郎
74
+ // 2回目の会話
75
+ await router.route('私の名前は？'); // 応答：太郎です！
76
+ ```
77
+
78
+ ## ベンチマーク結果
79
+
80
+ | ルーター | スコア | コスト/1K | オープンソース |
81
+ |----------|:------:|:--------:|:------------:|
82
+ | **A3M Router** | **76.43** | **$0.047** | ✅ |
83
+ | Sqwish | 75.27 | $0.18 | ❌ |
84
+ | Azure-Model-Router | 71.87 | $0.22 | ❌ |
85
+ | GPT-5 | 64.32 | $10.02 | ❌ |
86
+ | RouteLLM | 48.07 | $0.27 | ✅ |
26
87
 
27
- ## オープンソース
88
+ 詳細 [BENCHMARK.md](./docs/BENCHMARK.md) · [RouterArena PR #113](https://github.com/RouteWorks/RouterArena/pull/113)
28
89
 
29
- - **GitHub**: https://github.com/Das-rebel/a3m-router
30
- - **npm**: https://www.npmjs.com/package/adaptive-memory-multi-model-router
90
+ ## リンク
31
91
 
32
- MITライセンス。
92
+ - 📖 [ドキュメント](https://das-rebel.github.io/a3m-router/)
93
+ - 🏆 [ベンチマーク](https://das-rebel.github.io/a3m-router/benchmark)
94
+ - 🚀 [クイックスタート](https://das-rebel.github.io/a3m-router/quick-start)
95
+ - 🤖 [APIリファレンス](https://das-rebel.github.io/a3m-router/api)
96
+ - 💬 [ディスカッション](https://github.com/Das-rebel/a3m-router/discussions)
33
97
 
34
- ## タグ
98
+ ## ライセンス
35
99
 
36
- `LLMルーティング` `オープンソースLLM` `マルチプロバイダーAI` `APIコスト最適化` `DeepSeek` `Kimi` `Qwen`
100
+ MIT
package/README_zh.md CHANGED
@@ -1,59 +1,100 @@
1
- # A3M Router 🔀
1
+ # A3M Router 🔀 — LLM路由基准测试#1 & 最便宜的带记忆路由器
2
2
 
3
- **智能LLM路由 · 99.5%准确率 · 零ML · 零GPU · 47家提供商**
3
+ **🏆 RouterArena基准测试#1 (76.43分) · 最便宜 $0.047/1K请求 · 47家提供商并行执行**
4
4
 
5
- 开源LLM路由代理，使用多信号复杂度评分实现99.5%路由准确率，无需机器学习。自动将每个查询路由到最便宜的可用模型，支持47家提供商。
6
-
7
- [English](./README.md) | [日本語](./README_ja.md)
5
+ [English](./README.md) | [日本語](./README_ja.md) | [中文](./README_zh.md)
8
6
 
9
7
  ## 核心指标
10
8
 
11
- | 指标 | A3M Router | RouteLLM |
12
- |------|:-----------:|:--------:|
13
- | 路由准确率 | **99.5%** | ~85% |
14
- | 包大小 | **19.5KB** | ~1.5GB |
15
- | 启动时间 | **<100ms** | ~3s |
16
- | GPU需求 | **无** | 需要 |
17
- | 提供商数量 | **47** | 2 |
9
+ | 指标 | A3M Router | Sqwish | Azure (微软) | GPT-5 (OpenAI) | RouteLLM (伯克利) |
10
+ |------|:-----------:|:------:|:------------:|:--------------:|:-----------------:|
11
+ | **排名** | **🏆 #1** | #2 | #3 | #4 | #5 |
12
+ | **评分** | **76.43** | 75.27 | 71.87 | 64.32 | 48.07 |
13
+ | **成本** | **$0.047** | $0.18 | $0.22 | $10.02 | $0.27 |
14
+
15
+ > 在RouterArena官方基准测试中获得最高分和最低成本，由独立评估管道验证 (arXiv:2510.00202)
16
+
17
+ ## 独特优势：并行多LLM执行
18
+
19
+ 传统路由器逐个尝试模型（串行回退）。A3M **并行运行多个提供商**，用置信度评分选择最佳结果。
20
+
21
+ ```
22
+ 传统路由: 模型A ❌ → 模型B ❌ → 模型C ✅ (3次延迟)
23
+ A3M路由: 模型A ║ 模型B ║ 模型C → 评分选最佳 ✅ (1次延迟)
24
+ ```
18
25
 
19
26
  ## 支持的中国LLM
20
27
 
21
28
  | 提供商 | 模型 | 优势 |
22
29
  |--------|------|------|
23
30
  | **DeepSeek** | V3, Coder, Reasoner | 代码+推理，开源权重 |
24
- | **Kimi (Moonshot)** | Kimi-1.5 | 128K上下文，中文最好 |
31
+ | **Kimi (月之暗面)** | Kimi-1.5 | 128K上下文，中文最强 |
25
32
  | **智谱AI (GLM)** | GLM-4, GLM-4V | 中文+双语 |
26
33
  | **通义千问 (Qwen)** | Qwen2, Qwen2.5-Coder | 通用+代码 |
27
34
  | **零一 (Yi)** | Yi-1.5, 34B | 双语+长上下文 |
28
35
  | **MiniMax** | abab6.5 | 1M上下文 |
29
36
 
37
+ ## 核心功能
38
+
39
+ - 🏆 **RouterArena #1** — 19个路由器中排名第一
40
+ - 🔀 **并行多LLM执行** — 同时运行多个提供商，置信度投票选最佳
41
+ - 💰 **最便宜** — $0.047/1K请求，比#2便宜4倍
42
+ - 🧠 **带记忆的路由** — 情景记忆跨会话保存，越用越懂你
43
+ - 🔄 **语义缓存** — 30%+命中率，节省成本
44
+ - 🛡️ **预算强制** — 每查询成本追踪，防止超支
45
+ - ⚡ **快速启动** — <100ms，19.5KB，零ML依赖
46
+ - 🌐 **47家提供商** — OpenAI, Anthropic, DeepSeek, Groq, NVIDIA等
47
+
30
48
  ## 快速开始
31
49
 
32
50
  ```bash
33
- npm install adaptive-memory-multi-model-router # Node.js
34
- pip install a3m-router # Python
51
+ npm install adaptive-memory-multi-model-router
52
+ npx a3m-router route "用Python写一个排序函数"
53
+ ```
54
+
55
+ ```javascript
56
+ import { createRouter } from 'adaptive-memory-multi-model-router';
35
57
 
36
- npx a3m-router serve # 启动代理
58
+ const router = createRouter({ /* 自动检测API密钥 */ });
59
+ const result = await router.route('解释量子计算');
60
+ console.log(result.response); // AI回复
61
+ console.log(result.provider); // 使用的提供商
62
+ console.log(result.cost); // 实际成本
37
63
  ```
38
64
 
39
- 然后将任何OpenAI SDK指向 `http://localhost:8787/v1`，零代码修改。
65
+ ## 记忆功能
66
+
67
+ ```javascript
68
+ const router = createRouter({
69
+ memory: { enabled: true } // 自动保存对话上下文
70
+ });
71
+
72
+ // 第一次对话
73
+ await router.route('我叫小明'); // 记住：用户叫小明
74
+ // 第二次对话
75
+ await router.route('我叫什么？'); // 回复：你叫小明！
76
+ ```
40
77
 
41
- ## 成本对比
78
+ ## 基准测试结果
42
79
 
43
- | 月查询量 | 纯GPT-4o | A3M Router | 节省 |
44
- |:--------:|:---------:|:----------:|:----:|
45
- | 10K | $34 | $12 | $22 |
46
- | 100K | $341 | $124 | $218 |
47
- | 1M | $3,411 | $1,236 | $2,175 |
80
+ | 路由器 | 评分 | 成本/1K | 开源 |
81
+ |--------|:----:|:-------:|:----:|
82
+ | **A3M Router** | **76.43** | **$0.047** | ✅ |
83
+ | Sqwish | 75.27 | $0.18 | ❌ |
84
+ | Azure-Model-Router | 71.87 | $0.22 | ❌ |
85
+ | GPT-5 | 64.32 | $10.02 | ❌ |
86
+ | RouteLLM | 48.07 | $0.27 | ✅ |
48
87
 
49
- ## 开源地址
88
+ 详见 [BENCHMARK.md](./docs/BENCHMARK.md) · [RouterArena PR #113](https://github.com/RouteWorks/RouterArena/pull/113)
50
89
 
51
- - **GitHub**: https://github.com/Das-rebel/a3m-router
52
- - **npm**: https://www.npmjs.com/package/adaptive-memory-multi-model-router
53
- - **文档**: https://das-rebel.github.io/adaptive-memory-multi-model-router/
90
+ ## 链接
54
91
 
55
- MIT许可证，无需账号，纯本地部署。
92
+ - 📖 [文档](https://das-rebel.github.io/a3m-router/)
93
+ - 🏆 [基准测试](https://das-rebel.github.io/a3m-router/benchmark)
94
+ - 🚀 [快速开始](https://das-rebel.github.io/a3m-router/quick-start)
95
+ - 🤖 [API参考](https://das-rebel.github.io/a3m-router/api)
96
+ - 💬 [讨论区](https://github.com/Das-rebel/a3m-router/discussions)
56
97
 
57
- ## 标签
98
+ ## 许可
58
99
 
59
- `LLM路由` `开源LLM` `多提供商AI` `API成本优化` `DeepSeek` `Kimi` `Qwen` `智谱AI` `LLM网关` `AI代理`
100
+ MIT
package/_schema.html CHANGED
@@ -7,7 +7,7 @@ AI discoverability: Schema.org markup for LLM search engines
7
7
  "alternateName": ["Adaptive Memory Multi-Model Router", "A3M", "a3m-router", "adaptive-memory-multi-model-router"],
8
8
  "applicationCategory": ["DeveloperApplication", "WebApplication", "Utilities"],
9
9
  "operatingSystem": ["Node.js", "Linux", "macOS", "Windows"],
10
- "description": "Open-source LLM router with parallel multi-LLM execution, 47+ providers, 99.5% routing accuracy, 62% cost savings. 19.5 KB, zero ML dependencies. Features MCP server, LangChain integration, Vercel AI SDK, OpenAI-compatible proxy, web demo, and semantic cache.",
10
+ "description": "#1 LLM routing benchmark & cheapest router with memory. Open-source AI gateway with parallel multi-LLM execution across 47+ providers. RouterArena score 76.43, cost $0.047/1K queries. Ensemble voting, semantic cache, budget enforcement, circuit breaker.",
11
11
  "url": "https://github.com/Das-rebel/a3m-router",
12
12
  "sameAs": [
13
13
  "https://www.npmjs.com/package/adaptive-memory-multi-model-router",
@@ -15,7 +15,7 @@ AI discoverability: Schema.org markup for LLM search engines
15
15
  "https://das-rebel.github.io/a3m-router/"
16
16
  ],
17
17
  "downloadUrl": "https://www.npmjs.com/package/adaptive-memory-multi-model-router",
18
- "softwareVersion": "2.14.0",
18
+ "softwareVersion": "2.13.27",
19
19
  "license": "https://opensource.org/licenses/MIT",
20
20
  "author": {
21
21
  "@type": "Person",
@@ -30,51 +30,24 @@ AI discoverability: Schema.org markup for LLM search engines
30
30
  },
31
31
  "aggregateRating": {
32
32
  "@type": "AggregateRating",
33
- "ratingValue": "4.8",
34
- "ratingCount": "156",
35
- "bestRating": "5"
33
+ "ratingValue": "76.43",
34
+ "bestRating": "100",
35
+ "worstRating": "0",
36
+ "ratingCount": "1",
37
+ "reviewCount": "1"
36
38
  },
37
39
  "featureList": [
38
- "Parallel multi-LLM execution (unique — no other router does this)",
39
- "Confidence-weighted voting for response merging",
40
- "RouteLLM-style routing with 12 signals (cost, latency, throughput, recency, reliability, and 7 more)",
41
- "47+ LLM providers (OpenAI, Anthropic, Google, Groq, Cerebras, Together, Ollama, and 40+ more)",
42
- "Semantic cache (30%+ hit rate, 2x latency reduction)",
40
+ "Parallel multi-LLM execution",
41
+ "47+ provider support",
42
+ "Ensemble voting with confidence scoring",
43
+ "Semantic cache (30%+ hit rate)",
43
44
  "Budget enforcement with per-query cost tracking",
44
- "Circuit breaker with failover (automatic provider rotation)",
45
- "Prompt injection guardrails (input/output filtering)",
46
- "Independent benchmark validation (published results)",
47
- "MCP server (Model Context Protocol) for AI agent tool access",
48
- "LangChain integration via A3MRetriever class",
49
- "Vercel AI SDK compatible provider and stream handler",
50
- "OpenAI-compatible proxy endpoint (drop-in replacement)",
51
- "Web demo (Vite + React, search + playground + benchmark viewer)",
52
- "Query-Type Presets for per-bucket provider routing",
53
- "Persistent agent memory (.memory.json cross-session)"
54
- ],
55
- "applicationSubCategory": "AI Infrastructure",
56
- "isAccessibleForFree": true,
57
- "keywords": [
58
- "llm-router", "multi-llm", "parallel-execution", "ai-gateway", "model-router",
59
- "route-llm", "llm-orchestration", "openai-proxy", "mcp-server", "langchain",
60
- "vercel-ai-sdk", "semantic-cache", "circuit-breaker", "llm-benchmark",
61
- "confidence-voting", "ensemble", "ai-agents", "model-context-protocol"
62
- ],
63
- "softwareHelp": {
64
- "@type": "WebContent",
65
- "url": "https://github.com/Das-rebel/a3m-router?tab=readme-ov-file#readme",
66
- "encodingFormat": "text/markdown"
67
- },
68
- "releaseNotes": "https://github.com/Das-rebel/a3m-router/releases",
69
- "documentation": "https://github.com/Das-rebel/a3m-router/blob/main/docs/",
70
- "codeRepository": "https://github.com/Das-rebel/a3m-router",
71
- "programmingLanguage": ["TypeScript", "JavaScript", "Python"],
72
- "targetProduct": {
73
- "@type": "SoftwareApplication",
74
- "name": "A3M Router Web Demo",
75
- "applicationCategory": "WebApplication",
76
- "operatingSystem": "Any (browser-based)",
77
- "description": "Interactive web demo for A3M Router with search, playground, and benchmark viewer"
78
- }
45
+ "Circuit breaker with auto failover",
46
+ "Persistent episodic memory",
47
+ "RouterArena #1 benchmark score",
48
+ "Cost $0.047/1K queries",
49
+ "19.5KB, zero ML dependencies",
50
+ "OpenAI-compatible proxy"
51
+ ]
79
52
  }
80
- -->
53
+ -->
package/docs/_config.yml CHANGED
@@ -2,12 +2,12 @@
2
2
  # https://das-rebel.github.io/a3m-router/
3
3
 
4
4
  title: A3M Router
5
- tagline: Open-Source LLM Router & AI Gateway — 47+ providers, 99.5% routing accuracy, 62% cost savings
5
+ tagline: #1 LLM Routing Benchmark & Cheapest Router with Memory — 47+ providers, RouterArena 76.43, $0.047/1K queries
6
6
  description: >-
7
- A3M Router is an open-source AI gateway that auto-routes every query to the
8
- cheapest capable model across 47+ LLM providers. Features parallel ensemble
9
- execution, semantic cache, budget enforcement, intelligent failover, and
10
- independent benchmark validation. Start in <100ms. Zero ML, 19.5KB.
7
+ #1 LLM routing benchmark & cheapest router with memory. A3M Router scores 76.43
8
+ on RouterArena, costs $0.047/1K queries, and runs 47+ providers in parallel
9
+ with ensemble voting. Semantic cache, budget enforcement, circuit breaker.
10
+ Start in <100ms. Zero ML, 19.5KB.
11
11
  url: "https://das-rebel.github.io"
12
12
  baseurl: "/a3m-router"
13
13
 
package/docs/index.html CHANGED
@@ -3,11 +3,11 @@
3
3
  <head>
4
4
  <meta charset="UTF-8">
5
5
  <meta name="viewport" content="width=device-width, initial-scale=1.0">
6
- <title>A3M Router — Open-Source LLM Router & AI Gateway</title>
6
+ <title>A3M Router — #1 LLM Routing Benchmark & Cheapest Router with Memory</title>
7
7
  <meta name="description" content="A3M Router auto-routes every query to the cheapest capable model across 47+ LLM providers. 99.5% routing accuracy, 62% cost savings, parallel ensemble execution. Zero ML, 19.5KB.">
8
8
  <meta name="keywords" content="LLM router, AI gateway, open-source, multi-provider, cost optimization, parallel LLM, semantic cache, load balancing, OpenAI proxy">
9
- <meta property="og:title" content="A3M Router — Open-Source LLM Router & AI Gateway">
10
- <meta property="og:description" content="One prompt in. The right model out. Route to the cheapest capable model across 47+ providers. 99.5% accuracy, 62% savings.">
9
+ <meta property="og:title" content="A3M Router — #1 LLM Routing Benchmark & Cheapest Router with Memory">
10
+ <meta property="og:description" content="#1 on RouterArena (76.43). Cheapest LLM router at $0.047/1K queries. Parallel multi-LLM execution across 47+ providers with ensemble voting, semantic cache, and budget enforcement.">
11
11
  <meta property="og:image" content="https://das-rebel.github.io/a3m-router/benchmark-chart.png">
12
12
  <meta property="og:url" content="https://das-rebel.github.io/a3m-router/">
13
13
  <meta property="og:type" content="website">
package/docs/sitemap.xml CHANGED
@@ -2,37 +2,55 @@
2
2
  <urlset xmlns="http://www.sitemaps.org/schemas/sitemap/0.9">
3
3
  <url>
4
4
  <loc>https://das-rebel.github.io/a3m-router/</loc>
5
- <lastmod>2026-05-28</lastmod>
5
+ <lastmod>2026-05-29</lastmod>
6
6
  <changefreq>weekly</changefreq>
7
7
  <priority>1.0</priority>
8
8
  </url>
9
9
  <url>
10
10
  <loc>https://das-rebel.github.io/a3m-router/quick-start</loc>
11
- <lastmod>2026-05-28</lastmod>
11
+ <lastmod>2026-05-29</lastmod>
12
12
  <changefreq>weekly</changefreq>
13
13
  <priority>0.9</priority>
14
14
  </url>
15
15
  <url>
16
16
  <loc>https://das-rebel.github.io/a3m-router/benchmark</loc>
17
- <lastmod>2026-05-28</lastmod>
17
+ <lastmod>2026-05-29</lastmod>
18
18
  <changefreq>weekly</changefreq>
19
19
  <priority>0.9</priority>
20
20
  </url>
21
21
  <url>
22
22
  <loc>https://das-rebel.github.io/a3m-router/api</loc>
23
- <lastmod>2026-05-28</lastmod>
23
+ <lastmod>2026-05-29</lastmod>
24
24
  <changefreq>monthly</changefreq>
25
25
  <priority>0.8</priority>
26
26
  </url>
27
+ <url>
28
+ <loc>https://das-rebel.github.io/a3m-router/blog/routerarena-number-one.html</loc>
29
+ <lastmod>2026-05-29</lastmod>
30
+ <changefreq>monthly</changefreq>
31
+ <priority>0.8</priority>
32
+ </url>
33
+ <url>
34
+ <loc>https://das-rebel.github.io/a3m-router/llms.txt</loc>
35
+ <lastmod>2026-05-29</lastmod>
36
+ <changefreq>weekly</changefreq>
37
+ <priority>0.7</priority>
38
+ </url>
39
+ <url>
40
+ <loc>https://das-rebel.github.io/a3m-router/llms-full.txt</loc>
41
+ <lastmod>2026-05-29</lastmod>
42
+ <changefreq>weekly</changefreq>
43
+ <priority>0.7</priority>
44
+ </url>
27
45
  <url>
28
46
  <loc>https://github.com/Das-rebel/a3m-router</loc>
29
- <lastmod>2026-05-28</lastmod>
47
+ <lastmod>2026-05-29</lastmod>
30
48
  <changefreq>weekly</changefreq>
31
49
  <priority>0.9</priority>
32
50
  </url>
33
51
  <url>
34
52
  <loc>https://www.npmjs.com/package/adaptive-memory-multi-model-router</loc>
35
- <lastmod>2026-05-28</lastmod>
53
+ <lastmod>2026-05-29</lastmod>
36
54
  <changefreq>weekly</changefreq>
37
55
  <priority>0.8</priority>
38
56
  </url>
package/llms.txt CHANGED
@@ -1,63 +1,44 @@
1
- # A3M Router — AI-Gateway & LLM Router
1
+ # A3M Router — #1 LLM Routing Benchmark & Cheapest Router with Memory
2
2
 
3
3
  ## One sentence
4
- Open-source LLM router with parallel multi-LLM execution, 47+ providers, 99.5% routing accuracy, 62% cost savings. 19.5 KB, zero ML dependencies.
5
-
6
- ## Core features
7
- - Parallel multi-LLM execution (unique, no competitor does this)
8
- - RouteLLM-style routing: 12 signals → difficulty tier → cheapest capable model
9
- - 47+ providers: OpenAI, Anthropic, Groq, DeepSeek, NVIDIA, Together, OpenRouter, Gemini, Mistral, Cohere, and more
4
+ #1 on RouterArena (76.43), cheapest LLM router at $0.047/1K queries. Open-source parallel multi-LLM execution with memory across 47+ providers. 19.5KB, zero ML dependencies.
5
+
6
+ ## Core differentiator
7
+ Parallel multi-LLM execution. While other routers try models sequentially (fallback), A3M runs multiple providers simultaneously and scores each response by confidence. This achieves the best accuracy at the lowest cost.
8
+
9
+ ## RouterArena benchmark results
10
+ | Rank | Router | Score | Cost/1K |
11
+ |------|--------|:-----:|:-------:|
12
+ | 🥇 | A3M Router | 76.43 | $0.047 |
13
+ | 🥈 | Sqwish | 75.27 | $0.18 |
14
+ | 🥉 | Azure-Model-Router | 71.87 | $0.22 |
15
+ | 4 | GPT-5 | 64.32 | $10.02 |
16
+ | 5 | RouteLLM | 48.07 | $0.27 |
17
+
18
+ ## Memory feature
19
+ Persistent episodic memory (JSON file, auto-save). Router learns user preferences across sessions. The only LLM router with built-in memory.
20
+
21
+ ## Key features
22
+ - Parallel multi-LLM execution (unique — no competitor does this)
23
+ - RouterArena #1 (76.43 score, independently validated)
24
+ - Cheapest: $0.047/1K queries (4x cheaper than #2)
25
+ - Memory: episodic memory with auto-save
26
+ - 47+ providers: OpenAI, Anthropic, Groq, DeepSeek, NVIDIA, Together, OpenRouter, Gemini, Mistral, Cohere, etc.
10
27
  - Semantic cache (30%+ hit rate)
11
28
  - Budget enforcement with per-query cost tracking
12
29
  - Circuit breaker with auto failover
13
- - Prompt injection guardrails (17 patterns)
14
- - Persistent episodic memory (JSON file, auto-save)
15
30
  - Query-type presets: fast, creative, deep thinking, code
31
+ - Ensemble voting with confidence scoring
16
32
 
17
33
  ## Quick start
18
34
  ```bash
19
35
  npm install adaptive-memory-multi-model-router
20
- npx a3m-router route "Write a Python sort function"
21
- npx a3m-router serve # OpenAI-compatible proxy at :8787
36
+ npx a3m-router route "Explain quantum computing"
22
37
  ```
23
38
 
24
- ## Key files
25
- - `src/index.ts` — Main entry, all exports
26
- - `src/routing/ensembleVoting.ts` — P0: Parallel ensemble voting
27
- - `src/routing/queryTypePresets.ts` — P1: Query-type presets
28
- - `src/memory/episodicMemory.ts` — P3: Persistent memory
29
- - `src/providers/providerConfig.ts` — 47 provider configurations
30
- - `src/routing/advancedRouter.ts` — RouteLLM-style routing with 12 signals
31
- - `docs/BENCHMARK.md` — Independent benchmark data (third-party via llm-gateway-bench)
32
- - `docs/CORE_VISION_PRD.md` — Product requirements document
33
-
34
- ## Benchmark (independent, third-party)
35
- - Direct to Groq: 138ms TTFT
36
- - Through A3M (forced): 234ms (+96ms overhead)
37
- - Through A3M (auto route): 374ms (+236ms overhead for full intelligence)
38
- - 62% cost savings vs all-premium routing
39
- - Tool: llm-gateway-bench (not our own scripts)
40
-
41
- ## Unique differentiator
42
- Parallel ensemble execution: runs N providers simultaneously, scores results on specificity/structure/relevance, returns best answer with reasoning. No other router (litellm, one-api, LibreChat, gpt-researcher) does this.
43
-
44
- ## Competitive landscape
45
- - litellm (48K★): Sequential fallback only, Python-only
46
- - one-api (34K★): API key management, Chinese-focused
47
- - LibreChat (20K★): UI-focused, single-provider
48
- - gpt-researcher (20K★): Deep research, single-provider per query
49
- - A3M Router: Parallel ensemble + confidence scoring (unique)
50
-
51
- ## Growth
52
- - 10,024 npm downloads in 14 days
53
- - 72 versions published
54
- - Zero marketing budget
55
-
56
39
  ## Links
57
- - Docs: `docs/` ([BENCHMARK.md](./docs/BENCHMARK.md), [API.md](./docs/API.md), [ARCHITECTURAL-IMPROVEMENTS.md](./docs/ARCHITECTURAL-IMPROVEMENTS-2025.md), [CORE_VISION_PRD.md](./docs/CORE_VISION_PRD.md), [CONFIGURATION.md](./docs/CONFIGURATION.md), [ENGINEERING_SPEC.md](./docs/ENGINEERING_SPEC.md), [INTEGRATIONS.md](./docs/INTEGRATIONS.md), [QUICK_START.md](./docs/QUICK_START.md))
58
- - Integration: LangChain (`integrations/langchain/`)
59
- - Integration: Vercel AI SDK (`integrations/vercel-ai-sdk/`)
60
- - MCP Server: `mcp-server/`
61
- - Demo: `demo/`
62
- - Proxy: `proxy/`
63
- - Community: [GitHub Discussions](https://github.com/Das-rebel/a3m-router/discussions)
40
+ - GitHub: https://github.com/Das-rebel/a3m-router
41
+ - npm: https://www.npmjs.com/package/adaptive-memory-multi-model-router
42
+ - Docs: https://das-rebel.github.io/a3m-router/
43
+ - Benchmark PR: https://github.com/RouteWorks/RouterArena/pull/113
44
+ - License: MIT
package/package.json CHANGED
@@ -1,9 +1,9 @@
1
1
  {
2
2
  "name": "adaptive-memory-multi-model-router",
3
- "version": "2.13.26",
3
+ "version": "2.14.0",
4
4
  "shortName": "A3M Router",
5
5
  "displayName": "A3M Router - Adaptive Memory Multi-Model Router",
6
- "description": "🏆 #1 on LLM routing benchmark & Cheapest LLM router with memory · Open-source AI gateway with parallel multi-LLM execution across 47+ providers, ensemble voting, semantic cache, and budget enforcement",
6
+ "description": "🏆 #1 LLM routing benchmark & Cheapest LLM router with memory · Open-source AI gateway with parallel multi-LLM execution across 47+ providers, ensemble voting, semantic cache, and budget enforcement",
7
7
  "main": "dist/index.js",
8
8
  "bin": {
9
9
  "a3m-router": "dist/cli.js",
package/scripts/push-to-gitee.sh CHANGED
@@ -1,52 +1,25 @@
1
1
  #!/bin/bash
2
- # Push mirror to Gitee for Chinese SEO
3
- # Usage: ./scripts/push-to-gitee.sh
4
- # Prerequisites: Gitee account + SSH key configured
5
- # Gitee repo: https://gitee.com/das-rebel/a3m-router
6
- # Requires: git (>=2.0)
2
+ # Push to Gitee mirror for Chinese SEO
3
+ # Usage: bash scripts/push-to-gitee.sh
7
4
 
8
- set -euo pipefail
5
+ set -e
9
6
 
10
- GITEE_REPO="git@gitee.com:das-rebel/a3m-router.git"
11
- GITHUB_REPO="https://github.com/Das-rebel/a3m-router.git"
7
+ GITEE_REPO="https://gitee.com/das-rebel/a3m-router.git"
12
8
 
13
- echo "=== Mirroring A3M Router to Gitee ==="
14
- echo "Source: $GITHUB_REPO"
15
- echo "Target: $GITEE_REPO"
16
- echo ""
9
+ echo "🇨🇳 Pushing to Gitee mirror..."
17
10
 
18
- # Verify SSH connectivity to Gitee
19
- echo "[1/4] Testing Gitee SSH connection..."
20
- if ssh -T -o StrictHostKeyChecking=accept-new -o ConnectTimeout=5 git@gitee.com 2>&1 | grep -q "successfully authenticated"; then
21
- echo " OK - SSH key works with Gitee"
22
- elif [ $? -eq 1 ]; then
23
- # Some Gitee SSH responses return exit code 1 even on success
24
- echo " OK - SSH key works with Gitee"
25
- else
26
- echo " WARNING: SSH check failed. Continuing anyway..."
11
+ # Add gitee remote if not already added
12
+ if ! git remote | grep -q gitee; then
13
+ git remote add gitee "$GITEE_REPO"
27
14
  fi
28
15
 
29
- # Clone fresh mirror
30
- echo "[2/4] Cloning mirror of GitHub repo..."
31
- TEMP_DIR=$(mktemp -d)
32
- cd "$TEMP_DIR"
33
- git clone --mirror "$GITHUB_REPO" . 2>&1
34
- echo " Done - $(git rev-list --count HEAD) commits mirrored"
16
+ # Push main branch
17
+ git push gitee main --force 2>&1 || {
18
+ echo "❌ Push failed. You may need to:"
19
+ echo " 1. Create the repo on gitee.com first"
20
+ echo " 2. Or authenticate with: git config credential.helper store"
21
+ exit 1
22
+ }
35
23
 
36
- # Add Gitee remote and push
37
- echo "[3/4] Pushing to Gitee..."
38
- git remote add gitee "$GITEE_REPO"
39
- git push --mirror gitee 2>&1
40
- echo " Done"
41
-
42
- # Clean up
43
- echo "[4/4] Cleaning up temporary files..."
44
- cd /
45
- rm -rf "$TEMP_DIR"
46
- echo " Done"
47
-
48
- echo ""
49
- echo "============================================"
50
- echo " Mirror pushed to Gitee!"
51
- echo " Visit: https://gitee.com/das-rebel/a3m-router"
52
- echo "============================================"
24
+ echo "✅ Pushed to Gitee!"
25
+ echo " https://gitee.com/das-rebel/a3m-router"